2013 IEEE/RSJ International Conference on
Intelligent Robots and Systems (IROS)
November 3-7, 2013. Tokyo, Japan
An Analytical Method to Detect Collision between Cylinders Using
Dual Number Algebra*
Rajeevlochana G. Chittawadigi and Subir K. Saha
Abstract— Cylinder, a common geometric entity has a
discontinuity at the joining of cylindrical surface and circular-
disks. Hence, collision detection between two cylinders in space
is a difficult task and few have reported formulations to solve
it. In this paper, a novel analytical methodology is proposed to
detect collision or intersection between two cylinders. The
configuration, i.e., position and orientation, between the
cylinders was represented using the four Denavit-Hartenberg
(DH) parameters plus two extra parameters. Dual Number
Algebra was used to extract these six parameters. Tests
involved in collision detection between the cylinders were
between the lines and rectangles in a plane, thus considerably
simplifying the problem of collision detection. As an
illustration, an one-DOF arm modeled as a cylinder with
cylindrical shaped obstacles were modeled and tested for their
collisions. The results were validated with an analytical method
available in the literature and a commercial software.
I. INTRODUCTION
The term ‘cylinder’ typically refers to the volume bound
by a right-circular cylindrical surface and two circular disks
as end-caps. Along with spheres, cuboids, cones, etc.,
cylinders are one of the main geometric primitives used in
Computer Graphics, Computer Aided Geometric Design and
related areas. In applications related to Robotics, medical and
surgical simulations, simulation of various processes,
intersection of different geometric objects in the environment
needs to be determined for realistic simulation, which is
referred as ‘Collision Detection’.
Collision detection, by itself is a very large field and
researchers have proposed different formulations for
geometric objects of various shapes [1]. To simplify the
process, typically, geometric objects (shapes) are enclosed
inside simpler bounding volumes such as spheres, Axis
Aligned Bounding Box (AABB), Oriented Bounding Box
(OBB), etc. and a broad phase test is done between the
bounding volumes. If these volumes intersect, narrow phase
testing is done between the actual geometry (shape) [2].
Cylinder, though a very commonly used geometric entity, is
usually not considered as a bounding volume because
collision detection between two cylinders is reported [3] to be
expensive. Instead, a capsule (cylinders with hemi-spherical
ends) has been suggested to be used as bounding volumes.
Note that the latter tests are quite simple.
* This research was funded by BARC/ BRNS, Mumbai, India, under the
project, “Setting up of a Programme for Autonomous Robotics” at
IIT Delhi, India.
R. G. Chittawadigi is an M.S. (Research) student in the Department of
Mechanical Engineering, Indian Institute of Technology Delhi, New Delhi,
India (e-mail: rajeevlochan.iitd@gmail.com).
S. K. Saha is a professor in the Department of Mechanical Engineering,
Indian Institute of Technology Delhi, New Delhi, India (phone: +91 11
26591135; fax: +91 11 26582058; e-mail: saha@mech.iitd.ac.in).
978-1-4673-6358-7/13/$31.00 ©2013 IEEE 5353
Exact collision detection between two cylinders in the 3D
Cartesian space, either for broad phase or narrow phase
testing has been reported by few. Eberly [3] considered
Separating Axis Tests (SAT) to determine overlap of two
cylinders along certain directions (each of the cylinder axes, a
vector connecting the centers of the cylinders and vectors on
a plane perpendicular to each of cylinder axes and passing
through their respective centers). It also required optimization
of a function over a spherical surface. A method was
proposed by Ketchel and Larochelle [4] using line geometry
to detect if two cylinders intersect, by determining the
common normal between the cylinder axes and the points
where the normal intersects the axes. Based on the condition
of these points lying inside respective cylinders, different
tests were proposed. However, detailed logical derivation
was not reported. Choi [5] reported cylinder as a Composite
Quadric Model (CQM), where cylindrical surface (quadric) is
bounded by a circular edge and plane on either side,
collectively referred as boundary elements. As CQMs are
semi-algebraic entities, firstly, algebraic methods were used
to compute contact points between all possible pairs of
boundary elements. Secondly, the contact points were
verified if they lay on the actual CQMs. Biermann et al. [6]
used optimization techniques to determine the distance
between different entities of cylinders (cylindrical surface,
circular edge and circular face). A concept of Minkowski
Portal Refinement (MPR) was used in XenoCollide [7] to
represent convex shapes, including cylinder, using a support
function and collision between two cylinders was reported
using an iterative technique. XenoCollide has been
implemented in libccd [8], an open-source collision library
for convex bodies. Kodam et al. [9] used cylinder-cylinder
intersection in Discrete Element Method (DEM) to simulate
the dynamics of particulate systems. They recognized six
contact scenarios between bands (cylindrical surfaces), faces
(circular faces) and edges (circular edges) and gave analytical
expressions to determine the occurrence and type of contact.
However, the axisymmetric nature of cylinders was not
exploited by them. Guo et al. [10] reported limitations with
methodologies proposed by Kodam et al. [9] in certain
configurations of cylinders and proposed several additional
tests and modifications to be incorporated for accurate tests
between finite cylinders. The cylinder-cylinder collision
module in Teikitu Gaming System [11] also uses separating
axis tests but it is based on heuristic solutions and no formal
derivation is available in open literature. An analytical
formulation was reported in [12] to find the proximity of two
cylinders in 3D space. Each cylinder comprised of four
geometric primitives and shortest distance between five
combinations of these primitives was used to conclude if two
cylinders collide or not. However, four of these were solved
in closed form, whereas the fifth combination required
solution of an eighth-order polynomial equation.
 In this paper, a novel analytical methodology is proposed A. Dual Angle
to detect collision between two finite cylinders in 3D space. It The concept of dual angle is now introduced which is used to
exploits the axisymmetric nature of the cylinders to simplify represent the relative displacement and orientation between
the tests which has not been used, to the best knowledge of two lines in space [14]. The dual angle ( ), as illustrated in
the authors, in the reported literature. The reminder of the Fig. 1(b), is defined as
paper is arranged as follows: Concepts of Dual Number
Algebra and Denavit-Hartenberg (DH) [13] parameters are (8)
defined in Section II, which are used in Section III for the
where is the projected angle between the lines and is the
derivation of proposed methodology. As an illustration,
shortest distance between them. If and are considered
collision between a rotating arm and few static cylindrical
dual unit vectors representing two lines and is the dual unit
objects was tested in Section IV. The results were verified
vector along common normal, the dual angle between them is
with those obtained using the algorithm of [4] and the
determined using
collision detection module of Autodesk Inventor software.
cos . (9)
II. DUAL NUMBER ALGEBRA
sin . (10)
A dual number or a dual vector is represented as a sum of
a real part and a dual part. The latter begins with the dual where , , y and are real numbers and the dual angle is
entity ε, which is nilpotent, i.e., ε2 = 0 [14]. Dual Number
Algebra is often used in the field of displacement analysis, arctan2 sin , cos (11)
kinematic synthesis and dynamic analysis of spatial
arctan2 ,
mechanisms as it produces concise and compact notations
[15-16]. The same when extended to vectors is known as Using (8) and (11), the relative orientation ( ) and distance
Dual Vector Algebra, which was used by the authors in [17] between the lines ( ) can be determined. Note that, for
to derive an analytical methodology to extract DH parameters parallel lines, no unique common normal exists and a method
of a serial robot from its CAD model. Here, the concept is used in [17] should be followed.
improvised and extended to test collision between cylinders.
A dual number ( is represented by
P
p
(1) Q

where and are real scalar number and corresponding

dual number, respectively, whereas ε2 = 0. Similarly, a dual
vector ( is expressed as (a) Line as dual vector (b) Dual angle between two lines
T T
(2) Figure 1. Line representation using Dual Number Algebra

where is called the resultant vector and is called the B. DH Parameters Based Transformation
moment vector, which are related as
To represent a coordinate frame with respect to another,
(3) six coordinates or parameters are required, of which three
correspond to the relative position (e.g., position vector) and
In (3), is the position vector of any point P on the line remaining three for their relative orientation (e.g., Euler
shown in Fig. 1(a). Operations on dual vectors are performed angles). Melchiorri [19] reported that four Denavit-
similar to those on Cartesian vectors. For example, dot- and Hartenberg (DH) [13] parameters, typically used in geometric
cross-products of two dual vectors, also referred to as line modeling of serial robots, and additional two parameters can
dot- and cross-products [14], are defined by also be used as the required six parameters, referred here as
. . (4) six DH (SDH) parameters. As illustrated in Fig. 2, the four
IV IV
DH parameters between frames and ,
. . . relates a line ( IV ) relative to frame . The extra
(5) parameter c is required to locate the origin of frame
along IV and parameter φ to orient V (parallel to IV ) with
. The six DH parameters are described in Table I.
In the same way, the Euclidean norm of a dual vector is
given by TABLE I. DESCRIPTIONS OF SIX DH PARAMETERS

. Parameters Description
(6)
b Distance between and I along
where and are the Euclidean norms of the dual θ Angle between and II about I
I
vector and vector , respectively. Dual unit vector is a Distance between II and III along II
defined as a dual vector whose Euclidean norm is equal to
unity [18], i.e.,
α Angle between III and IV about III
c Distance between IV and V along IV
(7) φ Angle between V and along V

5354
 V V IV IV
,
IV
I II
,
V V
,
II V

III II
I II V
IV
I III III
IV
II
II III
III ,
I
,
III IV I
,

Figure 2. Six DH parameters between two coordinate frames

Figure 4. Six DH parameters: Sequence θ, b, α , c, a and φ
Given the homogeneous transformation matrix of
Frame 2 ( ) with respect to Frame 1 ( , which is cylinders are axisymmetric, θ and φ do not play any role in
denoted as , the common normal (i.e., II = IV ) deciding whether two cylinders intersect or not. Hence, only
between and is determined. The corresponding SDH four parameters, namely b, a, α and c, between the frames
parameters can then be found as three dual angles between I I I
and V V
are sufficient. For the sake of
dual vectors mentioned in Table II and by using (8-11). convenience, Cartesian frame I I I henceforth is referred
V V
as and as . Note that all vector equations in the
TABLE II. SIX DH PARAMETERS AS DUAL ANGLES
formulation are referred in these frames.
Parameters Dual Angle Dual Vectors
A. Methodology
θ and b II
and As an input to the proposed methodology, transformation
α and a II IV ( ) of with respect to is required. Further the radii
and
IV ( and ) and half of the heights of the cylinders, referred
φ and c and here as half-extents ( and ), are needed. Given two
cylinders, one can find the SDH parameters using the Dual
III. COLLISION BETWEEN CYLINDERS Number Algebra presented in Section II. To detect collision,
A novel analytical method is proposed here to detect intersection of cylinders of infinite length is tested first by
collision between two cylinders using the SDH parameters checking if the separation distance (a) is greater than the sum
given in Table I. Referring to Fig. 3, a coordinate frame ( ) of radii ( + ). If so, the cylinders do not intersect and there
is attached to the center of the cylinder , whose Z axis ( ) is no need of further tests. Else, further tests on finite
is along the axis of the cylinder. Similarly, is attached to cylinders need to be performed. For finite cylinders, the
cylinder . For any transformation between them, following strategies are undertaken:
corresponding SDH parameters are determined using (8-11). • If the cylinders are parallel (i.e., α = 0° or 180° and
V
b = 0, c ), as shown in Fig. 5(a), the sum of half-
, V
extents ( and ) is checked with axial distance
between the centers of the cylinders, i.e., the value of
I II
, c. If the sum is greater or equal to the value of c, the
V
II
cylinders intersect. If the sum is less, the two
I cylinders do not intersect.
I III
IV
II
IV • For non-parallel cylinders, the points on cylinder
III
axes where the common normal ( II ) intersects are
tested. If these points are within respective cylinders
III
, IV
(i.e., | | and | | ), then the two cylinders
intersect, as illustrated in Fig. 5(b). This is similar to
“On-On Test” reported in [4]. Here it is referred as
Figure 3. Six DH parameters: Sequence b, θ, a, α , c and φ Non Parallel Test (NPT).
If either or both common normal points is/are outside
In this formulation, a modified sequence of
the respective cylinders, and are projected onto
transformations is used. The angular parameter θ is moved to
the beginning of the series of transformations and distance plane resulting in two rectangles and ,
parameter a, is moved towards the end, before angle respectively, as shown in Fig. 6. A simple Separating
parameter φ. This is done to take the advantage of cylinders Axis Test [2] is performed between the rectangles
being axisymmetric and to facilitate projected rectangles tests and if they do not intersect (Fig. 6(a)), the cylinders
in the proposed formulation. Note that the transformation do not intersect and no further test is required. The
coordinates of the vertices of the projected rectangles
between and is still the same, as shown in Fig. 4. Since

5355
 B
Volume of
intersection
B
A B
A
B
A

(a) Volume of intersection (c) Cross-section at

(a) Parallel Test (b) Non Parallel Test

Plane
Figure 5. Finite cylinder tests

are computed as explained later. If the projected A

rectangles intersect, as shown in Fig. 6(b), the two
cylinders at hand may or may not intersect, which is A
determined by the Vertex Edge Test explained
below.
(b) Boundary conditions (Top-view) (d) Cross-section at

Figure 7. Boundary conditions of Vertex Edge Test

Plane
For the cross-section plane at a distance of u from as
shown in Fig. 8, the expressions of the coordinates of vertices
of in A , are
T
(13)
(a) Non-intersecting rectangles (b) Intersecting rectangles
T
(14)
Figure 6. Projection Test between rectangles T
(15)
B. Vertex Edge Test T
(16)
First, a hypothesis is proposed which will be verified where and are half-breadth and half-extent,
later. respectively, of as shown in Fig. 8(a). Similarly, vertices
Hypothesis: If two cylinders are colliding, a volume of of in and after transforming to are given by
intersection exists between them which can be determined
using numerical integration techniques [20]. In the proposed B

methodology, this step is avoided by considering only the

boundary conditions of the volume of intersection along .
At these boundary conditions, the cross-sections of the A
Frame B
and on a plane parallel to result in rectangles and
, which would be touching each other.
B
For example, in Fig. 7(a), the volume of intersection of
two intersecting cylinders is illustrated, for which the
boundary conditions at and from along in the A

top-view shown in Fig. 7(b). The cross-section of the Frame A

cylinders at boundary condition is shown in Fig. 7(c).
(a) Touching rectangles
Note that the two rectangles are touching. Similarly
rectangles are shown touching for boundary condition of
Fig. 7(d).
Proof: The homogeneous transformation matrix ( )
representing B with respect to A is determined as
1 0 0
0 cos sin sin (12)
0 sin cos cos (b) Geometry of (c) Geometry of
0 0 0 1
Figure 8. Rectangles and at a boundary condition

5356
 T
; (17) 2 (30)
T
; (18) 4 4 (31)
T
; (19) 2 2 2 2 (32)
T
; (20) 6 2 2 2
where , and are half-breadth and half-extent 4 4 4 4 (33)
of , respectively.
2 2 (34)
The condition for touching at the boundary exists if two
2 2
rectangles just touch each other. It can be established by
constraining a vertex of one rectangle to lie on an edge of the 2 2
other. This can be derived as algebraic equation in u. The
tests are done for 32 possible combinations of edges and Equation (29) can be solved using the analytical technique
vertices, of which two are illustrated here. Tests on remaining reported in [21]. Note that only two solutions of u are real
combinations are similar. and geometrically feasible. If one or both of the real solutions
of u lie(s) in the interval defined in (24), the Z coordinate of
1) Vertex on Edge vertex is determined to find if lies on the edge
The condition for vertex to lie on edge is and not beyond it. Hence collision is concluded.
obtained by equating the Z components of vertex and
Note that by substituting u = 0 and a = 0 in (13-20), the
in using (13) and (20), as illustrated in Fig. 8(a), i.e.,
coordinates of the projected rectangles and , introduced
(21) in Fig. 6 can be determined.
which is rearranged as B Frame B
(22) B A

Using the geometry of , Fig. 8(c), a quadratic equation in u Frame A

is formed, i.e., A

(23)
After solving (23), one or both of the real solutions of u are
Figure 9. Condition for vertex to lie on edge
tested to lie in the interval defined by
| min , max , (24) IV. APPLICATION
For value(s) of u lying in the interval of (24), the Y For broad phase collision tests between robot links and
component of is tested if it lies on the edge and not objects in the environment, Oriented Bounding Box (OBB)
beyond it. If so, it is concluded that the projected rectangles was used in [22], whereas Axis Aligned Bounding Box
and intersect and hence the cylinders and (AABB) was used in [23]. If they collide, actual geometric
intersect and tests for the remaining combinations of vertices shapes of the robot links are then tested for collision. As most
and edges are not required. of the robot links are not rectangular blocks but cylindrical in
shape, the broad phase tests are too conservative. Hence, to
2) Vertex on Edge overcome this, the links are considered as cylinders
The condition for vertex to lie on edge , as (bounding volume), as also reported in the literature, e.g., [4].
shown in Fig. 9, is obtained by equating Y components of
and using (14) and (20), i.e., In this paper, as a proof-of-concept, cylindrical shaped
link for one-DOF arm and cylindrical obstacles were
cos sin (25) modeled in Autodesk Inventor software, as shown in Fig. 10.
Which, unlike (21), has term of . Substituting f ≡ 1, g ≡ The proposed analytical method was implemented as an
cos and h ≡ sin in (25), a general form is addin module inside the software using Visual C#. The joint
given by was rotated programmatically and the collision tests between
the arm and the obstacles were determined using the
0 (26) proposed methodology. The tests were also performed using
the methodology in [4] and collision detection module of
From cylinders’ geometry, Fig. 8(b-c), two equations are
Autodesk Inventor. The results of the collision tests, i.e., the
(27) boolean (true or false) value, matched perfectly for each
increment in joint angle, thus, validating the proposed
(28) methodology. Note that Autodesk Inventor uses a 3rd party
Combining (26-28), a quartic equation in u is then obtained commercial library by D-Cubed [24] for collision detection.
as follows: No detailed mathematical formulation on D-Cubed is
available in open literature. On the other hand, the authors of
0 (29) [4] used MATLAB environment for the implementation of
where the coefficients of the equation are their algorithm.

5357

Joint Axis
Obstacle 1
Obstacle 2
Arm
Obstacle 3
Figure 10. One-DOF arm with obstacles
For the proposed methodology, the sequence or the flow
of the collision tests conducted between ‘Arm’ and
‘Obstacle 1’, for one complete rotation of the joint axis, with
an increment of 1° is reported in Fig. 11. Note that majority
of the rejection occurred in “Infinite Cylinders Tests”. During
the finite cylinder tests, the “Projection Test” also had high
percentage of rejection due to which very few number of
“Vertex Edge Test”, which are relatively expensive, were
carried out. Similar results were obtained for the tests with
the other two obstacles having early exit from the tests.
Arm and Obstacle1
100%
F
Infinite Cylinders Test 65% No Collision
T “Cylindrical object contact detection for use in discrete element
35% 0%
Non Parallel Test Parallel Test
T F T F
9% 26% 0%
Collision Projection Test Collision No Collision
T F
5.5% 20.5%
Vertex Edge Test No Collision
T F
5% 0.5%
Collision No Collision
Figure 11. Flow of collision tests between ‘Arm’ and ‘Obstacle 1’
In terms of efficiency, the proposed algorithm appears to
have performed best among the three algorithms compared
for the present task, as it took CPU 1 time of about 0.05-0.12
milliseconds (ms), whereas those using [4] and Autodesk
Inventor took about 0.17-1.8 ms and 2-16 ms, respectively.
Since the above computations were carried out in different
software platforms, the CPU times may not convey the
effectiveness of the theoretical formulations provided by
different authors. However, they do reflect the efficiencies
the way they were implemented by the authors. In future,
more rigorous tests will be carried out for more complex
tasks and using other cylinder-cylinder collision formulations
reported in the literature in order to establish the robustness
and efficiency of the proposed algorithm.
V. CONCLUSIONS
Cylinders are one of the most common geometric entities
that are used in various fields of visualization but there are
very few exact and accurate methods to detect collision or
intersection between them. In this paper, an analytical
methodology is proposed for the same. The methodology
1
System Configuration: Windows XP 32 bit Operating System,
Intel Core 2 Duo (3.0 GHz) processor and 4 GB RAM
relies on planar geometries, thereby, making the collision
tests fairly straightforward, unlike those reported in the
literature which are based on either 3-dimensional spatial
geometries or iterative procedure. Based on the CPU times of
three different algorithms used to test the collision of the task
shown in Fig. 10, the proposed algorithm has been proven to
be accurate and most efficient.
REFERENCES
[1] M. C. Lin and S. Gottschalk, “Collision detection between geometric
models: A survey,” in Proc. IMA Conference on Mathematics of
Surfaces, vol. 1, pp. 602-608, 1998.
[2] C. Ericson, Real-time collision detection, Morgan Kaufmann, 2005.
[3] D. Eberly, Intersection of Cylinders, Technical Report, Magic
Software, 2000.
[4] J. Ketchel and P. Larochelle, “Collision detection of cylindrical rigid
bodies for motion planning,” in Proc. IEEE International Conference
on Robotics and Automation, pp. 1530-1535, 2006.
[5] Y. K. Choi, “Collision detection for ellipsoids and other quadrics,”
Ph.D. dissertation, Dept. Comput. Sci., Univ. of Hong Kong, 2008.
[6] D. Biermann, R. Joliet and T. Michelitsch, “Fast distance computation
between cylinders for the design of mold temperature control
systems,” Technical Report, TU Dortmund University, 2008.
[7] G. Snethen, “Xenocollide: Complex collision made simple,” Game
Programming Gems, vol. 7, pp. 165-178, 2008.
[8] Libccd, http://libccd.danfis.cz, accessed in July 2013.
[9] M. Kodam, R. Bharadwaj, J. Curtis, B. Hancock and C. Wassgren,
method simulations. Part I–Contact detection algorithms,” Chemical
Engineering Science, vol. 65 no.22, pp. 5852-5862, 2010.
[10] Y. Guo, C. Wassgren, W. Ketterhagen, B. Hancock and J. Curtis,
0%
“Some computational considerations associated with discrete element
modeling of cylindrical particles,” Powder Technology, vol. 228, pp.
193-198, 2012.
[11] Teikitu Gaming System, http://www.andrewaye.com/, accessed in
Sept. 2012.
Legend
[12] R. A. Srivatsan and S. Bandyopadhyay, “An analytical formulation for
T = True finding the proximity of two arbitrary cylinders in space,” Proc.
F = False International conference on Advances in Robotics, 2013.
[13] J. Denavit and R. S. Hartenberg, “A Kinematic Notation for Lower-
pair Mechanisms Based on Matrices,” ASME Journal of Applied
Mechanisms, vol. 22, no. 2, pp. 215–221, 1955.
[14] I. S. Fischer, Dual-Number Methods in Kinematics, Statics and
Dynamics, CRC Press, 1998.
[15] S. Bandyopadhyay, “Analysis and Design of Spatial Manipulators: an
Exact Algebraic Approach using Dual Numbers and Symbolic
Computations,” Ph.D. dissertation, Dept. Mech. Eng., Indian Institute
of Science, Bangalore, India, 2006.
[16] E. Pennestrì and R. Stefanelli, “Linear Algebra and Numerical
Algorithms using Dual Numbers,” Multibody System Dynamics,
vol. 18, no. 3, pp. 323-344, 2007.
[17] C. G. Rajeevlochana, S. K. Saha and S. Kumar, “Automatic
Extraction of DH Parameters of Serial Manipulators using Line
Geometry,” in CD Proc. The 2nd Joint International Conference on
Multibody System Dynamics, 2012.
[18] M. A. González-Palacios and J. Angeles, Cam Synthesis, Springer,
1993.
[19] C. Melchiorri, “Kinematic Model of Robot Manipulators,” Online
Lecture Notes, University of Bologna, accessed in Sept. 2012.
[20] Y. T. Lee and A. A. Requicha, “Algorithms for computing the volume
and other integral properties of solids. I. known methods and open
issues,” Commun. of the ACM, vol. 25, no. 9, pp. 635-641, 1982.
[21] J. Schwarze, “Cubic and quartic roots," Graphics Gems, Academic
Press Professional, Inc., 1990.
[22] T. Reichenbach and Z. Kovacic, “Derivation of Kinematic Parameters
from a 3D Robot Model Used for Collision-free Path Planning,” in
11th Mediterranean Conference on Control and Automation, 2003.
[23] X. Yang, Y. Zhao, W. Wu and H. Wang, “Virtual Reality Based
Robotics Learning System,” in Proc. IEEE International Conference
on Automation and Logistics, pp. 859-864, 2008.
[24] D-Cubed, www.siemens.com/D-Cubed, accessed in Sept. 2012.
5358